Mean and std values for transforms.Normalize

vision
1

Hi,
How do I choose the values for mead and std when using transforms.Normalize(mean, std)?

I have seen both examples where Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])]) are used, but also cases where Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]) are used.
How are these values found; should they be calculated from my data set or are they appropriate constants?

An example from our codebase looks like this:

def get_transform_eval(input_shape):
    return transforms.Compose(
        [transforms.ToPILImage(),
         transforms.Resize(input_shape),
         transforms.ToTensor(),
         transforms.Normalize(MEAN, STD)])
2

@void32 the values taken into consideration for mean and std for the transforms:

Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]) is used for the ImageNet dataset

And,

Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])] is used to convert input image colors to gray scale
You can check Understanding transform.Normalize( ) for a better insight

3

Hi Hemant,
Thanks for your quick answer.
Okay, it’s simply the mean and std calculated from the ImageNet dataset, so in theory, I should calculate for my own data?

4

@void32 yes, for a custom dataset to get better results